PAARTH SHARMA
About
I am a curious mind. I like coming up with innovative ideas based on what I experience in my daily life and the things that I learn. Then, I try to think of ways that the human frontier could be pushed with these concepts. My goal in life is to reach a level where most of the work that I do on a daily basis is pushing the bounds of knowledge of humanity.
Experience
IlluminexAICanada
- ›Making software solutions for company products
WATonomousCanada
- ›Developing a CUDA C++ inference engine for Meta's V-JEPA 2-AC world model
- ›Built a C++ robot simulation with Foxglove visualization↗ watch
UW OrbitalCanada
- ›Developed an LSTM anomaly detection pipeline for satellite telemetry
- ›Built a real-time satellite telemetry dashboard (React + FastAPI)
Jan – Apr 26HostalkyCanada
- ›6-stage parallelized multi-agent LLM pipeline converting doctor-patient transcripts into structured clinical notes
- ›Results: 12% higher Token F1 · 75% reduction in hallucination · traceable citations
Nov 25 – Aug 26Sample JunctionCanada
- ›E2E LangGraph lead enrichment pipeline
- ›Processing 300+ client profiles monthly with Pinecone embeddings
- ›Saving the sales team 100+ hours of manual research
Education
University of Waterloo
B.A.Sc. Computer Engineering
2025 → 2030 · Waterloo, ON
Writing
2026project ↗2026blog ↗
view all writing →neoKernel: a research loop for faster inference
Self-improving harness for Qwen3 inference on H100. A model proposes a bounded kernel or engine change; the loop guards the source, runs correctness checks, benchmarks against native, and keeps or reverts the candidate. 27 trials, ~700-800 tok/s.
Training a 30M-param GPT from scratch
Multi-head self-attention, positional encoding, training on tinyStories with JAX / Optax / Orbax on a T4. What broke, what worked, what I'd do differently.